BijliCast / README.md
muhammadumar15's picture
Update README.md
2848c31 verified
|
Raw History Blame Contribute Delete
4.36 kB
metadata
title: BijliCast
emoji: ⚡
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 6.32.0
app_file: app.py
pinned: false
python_version: 3.12.12
short_description: Next-hour household electricity forecasting for Pakistan

⚡ BijliCast — next-hour household electricity forecast

Weather-aware short-term load forecasting for Pakistani households. The model predicts the next hour's electricity consumption (kWh) from the last 168 hours of smart-meter usage, current-hour weather, calendar features and the household profile. No next-hour weather is required, so the forecast does not depend on a perfect weather forecast.

This Space is the live demo of the CT-378 Machine Learning project. Source, notebooks and the full report: BijliCast on GitHub.

Try it

  1. Pick one of the 59 households (six Pakistani cities).
  2. Adjust the current-hour weather if you want a what-if scenario.
  3. Press Predict next hour.

You get the point forecast, a 90% prediction interval and the last 72 hours of actual usage.

Model

Item Value
Deployed model HistGradientBoosting (scikit-learn pipeline)
Target next_hour_energy_kwh
Training rows 408,048 (train + validation; test never used for fitting)
Features 42 raw → 49 after calendar/wind encoding, scaling and one-hot
Selection chronological 70/15/15 split, tuning on train only
Validation TimeSeriesSplit(n_splits=5, gap=120) + randomized search
Artifact models/deployment_model.joblib (~1.4 MB)

Held-out test results (evaluated once, after selection)

Model Validation MAE Test MAE Test RMSE Test R²
Random Forest (validation winner) 0.3216 0.2535 0.4322 0.7229
HistGradientBoosting (deployed) 0.3273 0.2499 0.4273 0.7291
Linear Regression 0.3328 0.2601 0.4595 0.6869
KNN 0.4024 0.3178 0.5196 0.5996
Persistence (lag_1h) 0.3448 0.2760 0.5470 0.5561
Seasonal naive (lag_24h) 0.4806 0.3522 0.6415 0.3897

Random Forest and HistGradientBoosting are statistically tied on validation (CV fold std ≈ 0.067). HistGradientBoosting is deployed because it matches RF's accuracy, has the best one-shot test MAE, and its artifact is ~1.4 MB instead of ~270 MB — small enough for GitHub and free hosting.

Prediction interval

The 90% interval uses the 5th/95th percentiles of out-of-sample validation residuals ([-0.782, +0.846] kWh), reaching 90.0% coverage on validation and 93.8% on test. The lower bound is clipped at zero.

Inputs

  • Household: 59 static profiles (city, people, rooms, appliances, area, …).
  • Usage history: the last 168 contiguous hourly values per house, bundled in models/usage_history.parquet.
  • Weather (current hour): temperature, humidity, dew point, precipitation, wind speed/direction, pressure, solar radiation. Defaults are the last known hour; sliders let you ask what-if questions.
  • The origin timestamp is used for calendar features; the target is origin + 1 h.

Data

  • LUMS REWD-P smart-meter dataset — 59 unique households, 515,786 hourly rows.
  • Open-Meteo ERA5 archive — six cities.
  • Household metadata from the REWD-P dataset.

The demo replays the last known week per household (meter data ends in Oct–Nov 2024); it is not a live meter feed.

Limitations

  • The test period (autumn) has a lower load level than validation (summer), so validation MAE is higher for every model. This is seasonal shift, not over-fitting; selection never used the test set.
  • One global model serves all 59 houses; no per-house models.
  • Inputs are household-level smart-meter readings and city weather; the model does not know appliance-level usage or outages.
  • Research/demo use only — not for billing or grid operations.

Files

app.py                     Streamlit entry point
streamlit_ui.py            UI (shared with the main repo)
service.py                 bundle loader + inference contract
models/                    HGB pipeline, profiles, 168 h history, metadata
src/                       minimal config / pipeline / inference modules
requirements.txt           pinned runtime